Nexlem Agents

The 16 registered agents of Nexlem v2.0: 14 pipeline agents (10 blog + 4 social channels) plus 2 standalone orchestrator tools that do not participate in any pipeline. Each agent is a Claude Code subagent: a markdown file in agents/<name>.md with a YAML frontmatter contract on top and a prompt body below.

All 16 agents are at the v1.0.0 baseline immediately after Phase 41 (the D-41-02 baseline). Individual agents bump per the strict D-41-03 policy: any change to an agents/*.md file requires a version bump in the same PR. See docs/CONFIG.md for the canonical frontmatter spec including the version: field and bump rules.

Channels

Agents live in three channels:

Channel	Count	Order Range	Purpose
blog	10	1–10	Produce the polished article and extract atoms
orchestrator	2	(no pipeline participation)	Standalone planning tools
social	4	(parallel after atomizer)	Derive social media assets from atoms

Pipeline Topology

For a full campaign type:

research (1)
  ↓
writer (2)
  ↓
fact-check (3)
  ↓
eeat (4)
  ↓
humanizer (5)
  ↓
seo (6)
  ↓
disclaimer (7)
  ↓
image-planner (8)
  ↓
quality-gate (9)            ← verdict gate; REJECTED halts the campaign here
  ↓
atomizer (10)               ← bridge: extracts atoms into atom_library
  ↓
  ├── carousel
  ├── caption                ← all four run in parallel
  ├── reel-script
  └── youtube-short

Campaign type variants:

Campaign Type	Agents Excluded from Pipeline
`full`	(none)
`blog-only`	atomizer, all social agents
`social-only`	research, writer, fact-check, eeat, humanizer, seo, disclaimer, image-planner, quality-gate (atomizer + social only)
`silo-cluster`	research (shared from pillar — D-12)
`blog-sibling`	research (shared from first site — D-17)
`refresh`	image-planner, atomizer, all social agents (content re-run only)

Blog Channel (10 agents)

research

Order: 1 Version: 1.0.0 Executor: real (Claude Code subagent) Accepts: topic, config.language, config.site_name (from campaign context) Produces: 01-research.json Tools: WebSearch, WebFetch, Write Campaign types: full, blog-only, refresh Purpose: Web research for the article topic. Emits competitor analysis, key facts, suggested outline, keywords list, and a minimum of 3 sources.

Output shape:

{
  "topic": "string",
  "sources": [{ "url": "string", "title": "string", "summary": "string" }],
  "competitor_analysis": "string",
  "key_facts": ["string"],
  "suggested_outline": ["string"],
  "keywords": ["string"]
}

Acceptance criteria: minimum 3 sources; topic, sources, competitor_analysis, key_facts, suggested_outline, keywords all present.

writer

Order: 2 Version: 1.0.0 Executor: real Accepts: 01-research.json, config.language, config.site_name, config.role, config.relationship Produces: 02-writer.md Tools: Read, Write Campaign types: full, blog-only, silo-cluster, blog-sibling, refresh Purpose: Writes the full article from the research. Enforces minimum 800 words, image-placeholder embedding (placeholder:img-N tags), and relationship voice.

Acceptance criteria: minimum 800 words; valid markdown; includes placeholder:img-N image tags; relationship voice enforced.

fact-check

Order: 3 Version: 1.0.0 Executor: real Accepts: 02-writer.md, 01-research.json Produces: 03-fact-check.md, 03-fact-check-report.json Tools: Read, Write, WebSearch Campaign types: full, blog-only, silo-cluster, blog-sibling, refresh Purpose: Verifies factual claims in the draft against research sources. Emits both a corrected article and a structured verification report.

Acceptance criteria: report JSON present; claims_verified field present.

eeat

Order: 4 Version: 1.0.0 Executor: real Accepts: 03-fact-check.md, 01-research.json Produces: 04-eeat.md, 04-eeat-meta.json Tools: Read, Write Campaign types: full, blog-only, silo-cluster, blog-sibling, refresh Purpose: E-E-A-T polish — strengthens Expertise, Experience, Authoritativeness, and Trust signals in the article.

Acceptance criteria: E-E-A-T signals present in output; meta JSON present.

humanizer

Order: 5 Version: 1.0.0 Executor: real Accepts: 04-eeat.md Produces: 05-humanizer.md Tools: Read, Write Campaign types: full, blog-only, silo-cluster, blog-sibling, refresh Purpose: Humanizes prose — removes AI-detectable patterns, enforces the configured relationship voice, and adds natural stylistic variation while preserving all facts and image placeholders.

Acceptance criteria: article rewritten; all facts preserved; all image placeholders preserved.

seo

Order: 6 Version: 1.0.0 Executor: real Accepts: 05-humanizer.md, 01-research.json Produces: 06-seo.md, 06-seo-meta.json Tools: Read, Write Campaign types: full, blog-only, silo-cluster, blog-sibling, refresh Purpose: SEO pass — rewrites headings, injects target keywords at appropriate density, generates meta title and description, and adds internal link suggestions.

Output shape (06-seo-meta.json):

{
  "title": "string",
  "description": "string",
  "target_keyword": "string",
  "slug": "string"
}

Acceptance criteria: meta JSON present; title and description present.

disclaimer

Order: 7 Version: 1.0.0 Executor: real Accepts: 06-seo.md Produces: 07-disclaimer.md Tools: Read, Write Campaign types: full, blog-only, silo-cluster, blog-sibling, refresh Purpose: Appends the brand disclaimer when the article matches disclaimer_required_for tags (see docs/CONFIG.md business.json). Original article content is always preserved.

Acceptance criteria: disclaimer present if required; original article content preserved.

image-planner

Order: 8 Version: 1.0.0 Executor: real Accepts: 07-disclaimer.md, config.visual_brand Produces: 08-image-planner.json Tools: Read, Write Campaign types: full, blog-only, silo-cluster, blog-sibling Purpose: Maps image placeholders in the article to alt-text and production briefs. Reads all placeholder:img-N tags and generates a detailed image-generation prompt for each, incorporating the visual_brand config. Excluded from refresh (refresh is content-only).

Output shape:

{
  "images": [
    {
      "id": "img-1",
      "prompt": "string",
      "alt": "string",
      "placement": "hero|inline"
    }
  ]
}

Acceptance criteria: prompts generated for all placeholders; visual_brand config applied.

quality-gate

Order: 9 Version: 1.0.0 Executor: real Accepts: 07-disclaimer.md, 04-eeat-meta.json, 06-seo-meta.json, 08-image-planner.json, config.quality_gates Produces: 09-quality-gate.json Tools: Read, Write Campaign types: full, blog-only, silo-cluster, blog-sibling, refresh Purpose: Scores the final article across 5 weighted criteria (clarity, E-E-A-T, originality, relevance, tone/voice) and issues an APPROVED or REJECTED verdict. Campaign halts here on rejection — social pipeline does not start.

Output shape:

{
  "verdict": "APPROVED|REJECTED",
  "overall_score": 0.0,
  "scores": {
    "clareza": 0.0,
    "eeat": 0.0,
    "originalidade": 0.0,
    "relevancia": 0.0,
    "tom_voz": 0.0
  },
  "feedback": "string"
}

Acceptance criteria: all score fields present; overall_score present.

atomizer

Order: 10 Version: 1.0.0 Executor: real (flipped from mock by Phase 38 ATOM-05) Accepts: 01-research.json, 07-disclaimer.md, 08-image-planner.json, config.social.enabled_channels Produces: 10-atomizer.json Tools: Read, Write Campaign types: full, blog-only, social-only, silo-cluster, blog-sibling Purpose: Extracts reusable content atoms (quotes, stats, definitions, examples, checklists) organized per social channel. Persists atoms with persist_to_library and atom_type flags into the atom_library table. Bridge into social fan-out — excluded from refresh (refresh is content-only).

Output shape:

{
  "carousel": { "hook": "string", "atoms": [...] },
  "captions": { "hook": "string", "atoms": [...] },
  "reel": { "hook": "string", "atoms": [...] },
  "youtube_short": { "hook": "string", "atoms": [...] }
}

Acceptance criteria: all enabled-channel keys present; each channel has a hook; atoms carry persist_to_library and atom_type flags.

Orchestrator Channel (2 agents — standalone)

These agents are NOT part of any pipeline. They are invoked by operator-driven commands and have campaign_types: [] in the registry.

silo-advisor

Version: 1.0.0 Executor: mock (intentional per D-40-08 — mock_allowed_in_uat: true) Accepts: 01-research.json Produces: 01b-silo-advisor.json Tools: Read, Write Campaign types: (none — standalone) Invoked by: /nexlem silo plan Purpose: Analyzes research and recommends single post vs silo, with a cluster plan when silo is recommended. Today's executor is a deterministic mock; future phases may flip to a real Claude Code subagent.

Output shape:

{
  "recommendation": "silo|single",
  "confidence": "high|medium|low",
  "content_density": "string",
  "topic_breadth": "string",
  "rationale": "string",
  "recommended_post_count": 0,
  "count_range": "string",
  "silo_plan": {}
}

keyword-expander

Version: 1.0.0 Executor: mock (intentional per D-40-08 — mock_allowed_in_uat: true) Accepts: seed_keyword, site_context (optional), count (optional, default 10, max 25) Produces: JSON array (stdout) Tools: Read, Write Campaign types: (none — standalone) Invoked by: /nexlem keywords expand Purpose: Expands a seed keyword into N related terms with rationale and relevance scores. LLM-only — no external paid APIs. Mock executor; same future-real path as silo-advisor.

Output shape:

[
  {
    "keyword": "string",
    "rationale": "string",
    "relevance": 0.85
  }
]

Acceptance criteria: valid JSON array; each item has keyword, rationale, relevance; relevance in [0, 1]; no external API calls.

All four run in parallel after the atomizer. Each consumes the atomizer output and the polished article. None depend on each other. All four were flipped from mock to real in Phase 39 (SOCIALV2-01).

carousel

Version: 1.0.0 Executor: real (flipped from mock by Phase 39 SOCIALV2-01) Accepts: 10-atomizer.json, social brand compliance config, social atom targets (carousel) Produces: 11-carousel.json Tools: Read, Write Campaign types: full, social-only Purpose: Generates an Instagram/LinkedIn carousel script — slide-by-slide content (headline, body, CTA) from atomizer output, applying brand voice and tone rules.

Output shape:

{
  "slides": [{ "headline": "string", "body": "string" }],
  "caption": "string",
  "hashtags": ["string"],
  "cta": "string"
}

Acceptance criteria: 3–10 slides; slides, caption, hashtags present.

caption

Version: 1.0.0 Executor: real (flipped from mock by Phase 39 SOCIALV2-01) Accepts: 10-atomizer.json, social hashtag config, social CTA rules (caption) Produces: 12-caption.json Tools: Read, Write Campaign types: full, social-only Purpose: Standalone post caption generator — produces platform-specific captions (Instagram, TikTok, YouTube) with per-platform character limits, hashtag caps, and CTA placement rules.

Output shape:

{
  "instagram": { "text": "string", "hashtags": ["string"], "cta": "string" },
  "tiktok": { "text": "string", "hashtags": ["string"], "cta": "string" },
  "youtube": { "text": "string", "hashtags": ["string"], "cta": "string" }
}

Acceptance criteria: all three platforms present; each has text, hashtags, cta.

reel-script

Version: 1.0.0 Executor: real (flipped from mock by Phase 39 SOCIALV2-01) Accepts: 10-atomizer.json, social CTA rules (reel) Produces: 13-reel-script.json Tools: Read, Write Campaign types: full, social-only Purpose: Short-form video script — hook-first 30–60 second script with scene breakdown, music mood, and CTA for reels and TikTok.

Output shape:

{
  "hook": "string",
  "scenes": [{ "duration_s": 0, "visual": "string", "narration": "string" }],
  "duration_target": 30,
  "music_mood": "string",
  "cta": "string"
}

Acceptance criteria: hook, scenes, duration_target, music_mood, cta all present.

youtube-short

Version: 1.0.0 Executor: real (flipped from mock by Phase 39 SOCIALV2-01) Accepts: 10-atomizer.json, social CTA rules (youtube_short) Produces: 14-youtube-short.json Tools: Read, Write Campaign types: full, social-only Purpose: YouTube Shorts script with explicit hook and payoff structure — narrated script with visual brief, optimized title, and description for YouTube's short-form format.

Output shape:

{
  "hook": "string",
  "scenes": [{ "duration_s": 0, "visual": "string", "narration": "string" }],
  "duration_target": 60,
  "title": "string",
  "description": "string",
  "cta": "string"
}

Acceptance criteria: hook, scenes, duration_target, title, description all present.

Frontmatter Spec

Every agent file's YAML frontmatter is validated against the FrontmatterSchema Zod 4 strict-object in src/lib/agent-frontmatter.ts. Required and optional fields are documented in docs/CONFIG.md under "Agent Prompt Frontmatter". The version: field (semver, required) was added in Phase 41 — all 16 agents start at 1.0.0.

Required fields: name, order, description, accepts, produces, acceptance_criteria, tools, version

Bump policy: Any change to an agents/*.md file — frontmatter or body, including typos — requires a version bump in the same PR. PATCH is the default; MINOR for additive changes; MAJOR for breaking output-shape changes. See docs/CONFIG.md for details.